/V916/_T/-/7 _8¢_

NASA-TM-7789619850026932

_ A Reproduced Copy

(% '{ !%",)Li

Reproduced for NASA

by the

NASA Scientific and Technical Information Facility

FFNo 672 Aug 65

https://ntrs.nasa.gov/search.jsp?R=19850026932 2020-06-03T10:02:17+00:00Z

_SA-TM-77896 1985

ULTRASONICINSPECTIONOFCARBONFIBERREINFORCEDPLASTICBYMEANSOFS/_PLE-RECOGNITIONMETHODSo(Ultt_dschallpruefung yon CFKMtttels MethodentierRustererkennung). RoBtlgramo AUgo1985o29po

lnte_att onal es Wehetech- May8-10,ntsches Symposiumueber 1984CFK--Etgenschaf_cen.Pruefung und_n_ yon KohlefaseeverstaerktenKunststoffen, Erdtng (WoGePo),po1-27

A85-16115(oPigo)

£ DISPLAY 05i-EJ !85N35E:A5e# ISSUE E:,q. PAGE 6076 CATEGORY E:_} RF"T#: NASA-TM-??896 NAS

!.!5:?7896 CNT#: NASW-_005 '_w=:...._i08iO0-EZ9PAGES UNCLASSIFIED DOCUMENT

( UTTL_ Uitvasc:'nici'ns'F..,eCtsa.v,,!_:le_reco!_.....,iti__.;,_eethodsir-_',,-.,of :-arbc:r',, fiber" rei'r',,'forced ._._'i._.o.:......-',,-,..,_,_..,._..by .....................,",.=, ....,"'-"_

(" CORF'_ Natio:.qa! Aer'or'_,:_uti,_-:s and Sp-,.ce Admir'_ist,,-'ati,::,r'..,.., Washir,,t:jton, r.,c°...... iAVAILoNT!S

E;,,-iF':!4C AOS.'iMF AO!

,Z:!....on" G.ERMANY,FEDEF;'.AL F;.°.EPL!F_:LIC OF T,..-a;,-,,s!. .--:::y Ka,.'.,-,,,,-_,-..-..:'_-,.':Leo) Associates, P,,.:..-:,dwc:,:::d fCi'l-:y- Ca!i!:",, O'rigi,..-..,ai ............;_,-,,- "; '--"._,-:_..... ,..,,, by Mess.,'er-schmitt-Boe!kow-B!ohm G.m.boH. •P,,-eser,,ted at !nte_-r',,° Weh',,-t._._,,=,,-_0,,Symp. ueb=,,,- ,_"_:"_,..,.,E!tg.s:,r'_scha-Fte'.,'",,, P_'-uei"-u.',,'-,,g ur,,d

Bev"e"-tu'ngS-J.OMay !9E..Vev°r" K°h!efase""v':'e'r'T,,-a,,-,,s].oi'ntoStaerENGLikter"SHo"i--coni--KL""nst_-"toi"-'f'e_-",,paper_ E"-di':'g""U].tr-ascha!Ipr'uef'u;,?g':'[.des'.).:: Ge,,"many_._. (.vc-,'r-,,CFK ,..,,,-,",t,.,_.,+',-_smeth,;::den de',,'- Mus'te',,-e'..,'-kennung :''_b!e,_=..--_.-e',,-sch.,:r_-.tt-Bc:elk:::,w-]-_-,.,.-.,h..m ,:.B,,m°b oH,, , Ic>8,'-i- p i-2?

MAJS.': ieL":ARBON FIBER REINFORCED PLA.ST!CSi-?-_-NONDESTRUCTIVE TESTSi.'*:'SAMPL!N";..":i/.i:.'.:-ULTRASONIC: FLAN DETErZ:TION

-'-'fINS: / ANALOG TO DIGITAL CONVERTERS/ COMPUTER GRAPH!CS/ COMPUTERS',/ E,:-:HOSOUNDINBi QUALITY CONTROL

ABA _ ......._ _',

,. ABS_ i'n the case oi = c--rbon f'iber" re-.r:i--o_-ced p].astic (CFRP)... it: has r_ot yet bee-,,p.,.?.ssib!e "to detect "nonlocal de'f.,.'-::,c'ks and mater-,ai deg.,."adat._.o'n ".,.---:!ated ......+_-,

ENTER: MORE

"+B!i!il........... - EM:_-'.--S-?

NASA-TECHNICALMH_D_! NASA TM-77896

ULTRASONIC INSPECTION OF CARBON FIBER REINFORCEDPLASTIC BY MEANS OF SAMPLE--RECOGNITION METHODS

R. Bilgram

Translation of "Ultraschallpr_fung yon CFK mittels Methodender Mustererkennung," Internationales Wehrtec_nisches

Symposium _ber CKF Eigenschaften, Pr_fung und Bewertungvon Kohlefaserverst_rkten Kunststoffen, Erding, West

Germany, May 8-10, 1984. Paper. pp. 1-27.

(NAS_-T_-77896) ULTRASONIC INSPECTION OF N85-352_5CARBON FIBE_ REINFOPCED PLASTIC BY MEANS OFSASPLE-R_COGNITION _ETHODS {NationalAeronautics and Space _dsini-stration) 29 p UnclasHC £_23/MFA01 CSCL 11D _3/2_ 27_55

NATIONAL AERONAUTICS AND SPACE ADMINISTRATIONWASHINGTON, D.C. 20546 AUGUST 1985

STANDARD TITLE PAGE

|, Report No. 2, Government Accession No. 3. RocJp_ent*s Catalog No,NASA-77896' "1'"ULTRASONICINSPEUrIONOF CARBONFIBER REINFORCED

PLASTICBY .MEANSOF SAMPLE-RECOGNITIONMEIIDDS 6. Per|ar,ingOrg.ni,o,ianCoda

7. Author(s) 8. Performing Organiaatlan Report No.

R. BILGRAM m. wo,kuoi,No.

- I% Controat or Grant No.9, psrfor_ng0,gon.ationN._..nJA_dres, NAS_/'4005

LEOKANNERASSOCIA]_S I'a'.'Typ.._Repo,,.,J Pori._C.... ,dREDWOODCITY,CA TRANSLATION

12. Sponsoring Agency Name and Add,el,NATIONALAERONAtrFICSANDSPACEADMINISTRATIONWASrtlNGTON,D.C. 20546 ,4. Sponso.rlogAga.cyCol, ,.

15. Supplementary Notes

Translation of "Ultraschallprueftmg van CFKmittels Methoden der Muster-erkmmung," Internationales Wehrtechnisches Symposiumueber CFKEige_schafon, Pruefung und Bewertung van Kohlenfaserverstaerkten Kunststoffen,Erding, West Germany, May 8-10, 1984, Paper. pp. 1-27(A85-16115 ,..

16. Abstract

In the case of carbon fiber reinforced _Iastic (CFRP), it has not yetbeen possible to detect non-local defects and material degradation relatedto aging with the aid of a nondestructive inspection • method. An approachfor overcoming the difficulties regarding ._uch an inspection involves anextension of the ultrasonic inspection procedure on the basis of a use ofsignal processing and sample recognition methods. The basic concept in-volved in this approach is related to the realization that the ultrasonicsignal contains information regarding the medium which is not utilizedin conventional ultrasonic inspection. IIowever, the analytical study of tl ephysical processes involved is very complex. For this reason, an empiricaapproach is employed to make use of the information which has not beenutilized before. This approach uses reference signals which can be ob-tained with material specimins of different qf_lity. The implementationof these concepts for the supersonic inspection of CFRPlarainates isdiscussed. ""

17. Key Words (Selected by Author(s)) 18. Distribution Statement

Unlimited Unclassified

19. Security Classif. (of this report) 20. Security Cla,sif. (of this pog*) 21. No. of Pages 22. Pric*_'"

Unclassified Unclassified 29

NASA.HQ

ULTRASONIC INSPECTION OF CARBON FIBER REINFORCEDPLASTIC BY MEANS OF SAMPLE RECOGNITION METHODS

R. BilgramMesserschmitt-Bolkow-Blohm GmbH

I. Introductio_ Z!.

Non-local defects such as manufacturing defects (e.g.

process irregularities, material flaws) and aging defects

(e.g. chemical, physical, mechanical degradation) in carbon fiber

reinforced plastics (CFRPs) have previously been impossible to

detect by nondestructive inspection methods. A set of solutions

is'seen in expanding in improving extensively proven ultrasound

inspection (US inspection) with methods of signal processing and

pattern recognition so that such defects can also be detected andmoreover a detailed defect type analysis and classification can

be performed during US inspection of a component.

The fundamental idea is that the US signal contains informa_'_

tion about the nature of the medium through wh[_chthe sound wave

has passed, or the kind of reflector from which it was reflected,which information (Fig_.'l)is not used in conventional US

inspection.

since an analytical treatment of physical processes during

sound propagation in components is extremely complex, and hence

practical inspection of the component is hardly helpful, an

empirical route is taken to make this information us_ful: A

measurement system is provided with reference signals such as are

measured on material samples-of_ar:ious qualities, with a signal

class assigned to each _qualityclass, suitably "trained," the

measurement system checks while measuring the component whether

it recognizes certain reference signal patterns in the signal,

.Numbers in the margin indicate pagination in the foreign text.1

and assignsthe patternto the correspondingsignalor qualityclass.

Below we discuss briefly how such a procedure can be

constructed, and what experience has been gained with it on

various CFRP laminates.

The method is founded on work by Rose et al. [1-7] and on

general literature on pattern-recognition [8]. L_

2. Method

2.1. General

The US wave train transmitted by the testing head into the

test specimen at the respective measurement point is reflected

from the interfaces in thespecimen accordinglto the laws of

sound propagation, is absorbed in the material and scattered by

microzcopic interfaces so that after a certain time the wave

train can be received with a correspondingly reduced sound

pressure as an echo or transmission signal (Fig. 2). In conven- ....

tional US inspection, the time and amplitude of the electronic-t.

ally smoothed and rectified received signal are evaluated, i.e.,

they are usually applied for "good/bad" judgments via a thres-

hold value discriminator within a time window. Interpretation is

thus limited to whether a sufficiently high echo signal isreceived Within the selected time window.

But precisely with media such as CFRPs, which strongly damp

sound, how the received signal is constituted after passing _-

through the material is a question of great interest -- i.e.,

what is the influence of various material quality features upon

the signal features. The consistent utilization of this connec-

tion requires application of known signal shape analysis ard

pattern recognition techniques, and _inally leads to an automatic

classification of US signals or qualities. The prerequisite is

that one must be able to verify (empirically) an interdependence

2

between signal features and material quality.

2.2. Measurement System

A measurement system for osucha patternrecognition method

in US inspection (Fig. 3) consists of the US transmitter/receiver

with the test head, coupled to the specimen manually via a

contact medium or in an immersion bath via a water separation /_3

space, and the analog/digital converter, which accepts the

received signals at a suitable scanning rate (up to 100 MHz) and

transmits them to the computer for further processing. The

results can be graphed or tabulated out via suitable peripher-

als. Via a monitor the US signal can be observed directly for

supervision.

The procedure for applying pattern recognition can roughly

be divided into three phases:

2.3. "Learninq" Phase

Reference signals for the various quality classes are

recorded from suitably prepared CFRP patterns ("good" signals and

various "defect" signals). From the signals, suitable signal <

features are calculated, e.g., rise time, pulse duration, decay

time of the envelope signal, various shape f_Gtors of the

envelope signal, the HF output spectrum, the so-called analytical

spectrum and/or the phase spectrum. Here "suitable" means:

physically meaningful and useful for a class differentiation. At

the end of the experiments, a data base is available containing a

suitable selection of signal feature values for each class to be

differentiated (e.g., 5), in a sufficient number for statistical

purposes (e.g., 10-20 or more measurement points for each class).

Using this data base, a computation rule (algorithm) is then

developed which using as few signal features as possible assigns

the measurement points of the data base to the right classes. In

principle this classification rule can be based on considerations

3

of tabulated values and establishment of limit values, but

preferably it uses graphlc/visual examination of population

diagrams a,_devaluation of probability densit_ functions of the

signal features of the classes being sought. Examples of this /__

are discussed further below. But since generally several classes

must be distinguished and up to 5 signal features must be

considered, human visual capacity or abstraction capability is

exceeded and one lets a computer develop algorithms according to

preset blueprints (i.e.,calculate the coefficients for so-calleddiscriminant functions), and then check the suitability of the

algorithm by inserting the values from the data base.

Here it must be noted that in this learning phase it is

possible to take account of the Scattering of various parametersfrom the areas of material characteristics and testing data,

which of course are reproducibl_ only within limits, by fully ....

including the corresponding scattering ranges in the base data.

2.4. Test Phase ....

Aspatterns of known error classes or quality features, by

measuring as many US signals as possible, deriving the feature

values and inserting these values in the algorithm from the

• learning phase, one checks how well the algorithm works at

arbitrary points of the known defect samples -- i.e., how high

._ the "hit rate" is. If it is not high enough, the learning phase

must be repeated, either expanding the data base with further

measurements, using other definitions for the signal features, or

if necessary altering the testing boundary conditions (testing

frequency, scanning window, amplification, etc.).

2.5. inspection-Phase

Given sufficiently high accuracy in the testingphase,

testing can now be performed with thedevelopedalgorithm on

unknown laminates. The result of each testing process is the

classification of the respective measurement_point, with a /__

4

certain probability < 100% (as with any non-destructlve test

result), which can be determined from the test phase.[

3. Sample Applications

2.1. Delamlnation Test

To illustrate the above description, a simple practical

example can be used, in which an algorithm was developed to

detect delaminations in a CFRP plate (5 mm thick, resin system914 C, fiber T 300). Figure 4 shows two US signals from this

example, as well as the HF spectrum, analytical spectrum, and

envelope of these signals. In the example only 2 classes, namelythe "good',range (Class I) and "delamlnatlon,,(Class 2) were to

be distinguished. Hence is was sufficient to consider the

relatively apparent features of the envelope, which are definedas shown in Fig. 5.

In the example in Fig. 6, one can see that feature No. 4 has

almost identical probability density functions for both classes,i.e., the same values are measured wlt_high probability for the

"good" range and delamination. Consequently the feature cannot

be valuable for classification. Feature No. 7 (upper half of

Fig. 6) on the other hand, separates the two classes better.

But as with the other features there is a relatively large rangeof overlap, meaning that a limit value that separates bc

glasses 100% does not exist for any feature. Consequently 2features must be used.

Here it is useful to consider the population graphs in

question, i.e., the values are graphed in a coordinate network as

in Fig. 7, formed by the feature axes. in the upper half, /_6features 5 and 6 were selected, and one sees that here it is

still impossible to separate Classes I and 2 100% by a line.

From the 45 possible Pairings of the I0 features, the pairing 6/7was found best. It is reproduced at the bottom of Fig. 7, and

rather than the centroid normal (solid line) drawn by the

5

computer, a line similar to the dotted line separates the twoclasses best.

The "blueprint" for the algorithm embodied in this separa-

tion line is a linear discriminant function with the general

form:

f = Co " ¢I Xl . c2 x2 ..... ¢n • Xn (I)

where xi are the feature values of the current measurement point,and c are the coefficients to be determined by the computer

program.

" The determined algorithm reads:

f = -0.0262- 3,9163 x6 + 1.8879 x7 (2)

The measurement values 6 and 7 determined at any given measure-

ment point (i.e., according to Fig. I0 these are the position of

the centroid of the envelope and the standard deviation of the

curve points) are inserted in Eq. (2) and the result f is

compared with a threshold value s = 0.782 and classified:

if f _ 0.782 Class 1 = "good"

i_ f < 0.782 Class 2 = "delamination"

A correspoDdingly structured test program supplies the classifi-

cation within ca. 4 sec., with an accuracy of nearly 100%. !_Z

3.2. Manufacturing Defects in Multidirectional Laminates

The actual domain of pattern recognition is US signals

that do not contain such obvious information as (e.g.) a delam-

ination-echo,-but-due-todeviating material-properties present

only certain signal alterations compared to a "good" standard.

These alterations may be due to deviations in process parameters

(pressure, temperature, time, etc.), contamination of the

laminates (moisture, mold lubricants_-separation films, etc.) and

aging effects (chemical, mechanical, physical).

6

JI

To give an example, plates of a 17-1ayer laminate type made

of unidirectional carbon fiber prepreg 914 C-/T300 were hardened

with various process defects. Figure 8 summarizes the 8

classes, i measurement point each was set on the 16 ,,good"

samples, and i0 measurement points each on the other samples, so

that in this case a data base of 86 measurement points is

available for the learning phase. The l0 features were defined

as in Fig. 5.

Here different studies were performed, i.e_ linear discrimi-

nant functions per Eq. (i) were determined for 2 classes each.

At this point we introduce another type of discriminant function,

the hyperquadratic, with the general form:

gi( )= + Vi. . Woi

with x = feature vector

xt = transposed feature vector

[Wi]= weight matrix

wi = weight vector _"

Wol = weight scalar J-_

For each class i, the value gi(x) is calculated; for the

class in question, gi(x) is the maximum.

Computer programs were developed which automatically

determine the best feature combinations for a given training data

set, i.e., the selection of signal features thatsupplies a

required minimum quota of correct class assignments for this dataset. In the first-step, an algorithm per Eq. (3) is developed

for each signal feature and each class, and the accuracy is

determined by inserting the training values. If the minimum

quota (90%)is reached, it is reported and the computer recom-

mends using_this_algorithm to solve the inspection problem.

Otherwise, in a second step all combinations of two are tested,7

then all combinations of three and so on, until a satisfactory

accuracy is reached. Experience shows that it is enough to

include 2 or 3 features, but for 8 classes sometimes 4 or more

features may be necessary.

An important questionis the quality of a classification

algorithm, i.e., what criterion should be used to evaluate which

of the possible algorithms or error finding allows the bestaccuracy.

For this one considers the so-called "performance matrix',in

Fig. 9 as a way of showing classification results in which the

number of correct class assignments for each class, the total

accuracy (= performance), detection quota (= detectability) and

"good" quota (= specificity) or Class 1 quota, are graphed.

_J

It is reasonable to expect such an algorithm primarily to

detect "good" points and "bad" points with optimum certainty,

and only secondarily to name the kind of defect correctly.

Therefore the mean of the quota for Class i (= "good") and /_

_ the error detection quota (mean of quotas for error classes) is

used as the evaluation criterion. Only when several possibil-

ities yield the same total q_ the total accuracy used as a secondcriterion.

In the present case, for example, the situation shown in

Fig. I0 developed. Using only one feature, features 5 and 8 each

yielded 87.8% as a mean criterion; the total accuracy of 37.7%

- here favors feature 5. With 2 features, the pairing of Nos. 12

and 19 with a mean criterion of 96.2% was the best, and here up

to 60.5% of all measured points could be assigned correctly.

The more features are considered, the higher the number of

right assignments, and the main question is how much computer

time per measurement pQint can be spent on classification.

8

The classification mechanism will be explained in more

detail with this example. As noted above, for each class a

discriminant function is-defined. InMFig. _I, 4 such functions

are graphed as an example for 4 classes, vs. the feature value.

Each measurement process yields a certain feature value and the

task of classification is to determine which of the functions at

this point has the highest value.

Even with 2 features, this classification is still manage-

able (Fig. 12). Here the discrlminant functions can be seen as

areas (with hills and valleys) intersecting at the classification

boundaries. Such intersecting lines are graphed in Fig. 12; the

numbers mean that at this point Of the "feature" plane the

respective function has the greatest value of all 8 functions./__

3.3_ Fabric Laminat_ _pecimens

While the sample plate studied in the above section _ad a uniform

wall thickness) in another test series, 27 specimens of 2

laminate types with 3 wall thicknesses were used, in a further

_ step towards close-to-practice testing problems. Laminate type A

is made purely of fabric prepregs (6, 12_a_d 18 layers). Type B

is made of alternating fabric layers and unidirectional tapes

(7, 13 and 19 layers). For each ofthe laminates, the correspon-

ding classification algorithms were determined in which i0 very

simple signal features (involving simple software) were used:

The first 5 features are defined like features 5-9 in

Fig. 5, applied to the rectified HF signal (software: f(t) =

/f(t)/). A further 5 features with the same definition were

drawn from the HF output spectrum....

Althoughhere we worked with greatly simplifiedsignal

features compared to the previous example, the defects were

distinguished very clearly (close to 100%)using only 2 or 3

9

features (Fig. 13).

There are two possible causes for the good classification

performance: First, the smaller number of classes (4 or 5), and

second the defects seem more severe than in the example with

multidirectional laminates (Sec. 3.3). /Ii

4. Summary and Prospects

The application of signal shape analysis and pattern

recognition methods opens new prospects for US inspection,

particularly of complex structures such as CFRPs. As was shown

with some simple laminate forms, signal features can be defined

and associated classification algorithms can be developed, with

which US signals can be assigned to certain reference signal

classes. Since these reference_._signalscome from laminates with

certain quality classes ("good" class and various defect clas-

ses), it is possible to assign US signals to a defect class.

The presented examples demonstrate that a measurement

system (measurement technology and software) can be constructed

to fulfil the posed task of detecting non-local production

_ defects such as process irregularities, contaminatibns, pretreat-

ment defects, etc., at detection rates of over 90%.

It can be assumed that certain aging defects can also be

detected. Studies on this will begin this year.

At the moment, work is aimed at discovering by steps the

requirements of a practical CFRP inspection, i.e. for detecting

real-defects-withthenecessary-certainty in real components in

an economically reasonable manner, even with more and more

complex component geometries and laminate types.

The next steps are:

i0

. -- Search for a test that is independent of wall thickness

and if possible of geometry. /12

-- Simplified operation and use of computer programs to

master many,sided inspection problems quickly.

-- Increased computation speed using Assembler subprograms.

-- Collecting further experience from previous results

concerning signal feature "quality," such as: Which ones respond

especially sensitively to certain types of defects, or how can

the "quality" of features best be evaluated.

-- Attempt at a physical explanation of the effect of

certain defects upon the US signals.

-- Study of possibilities for automatic defect detection,

without having reference specimens available.

-- Producing self-adapting changes in the classification

algorithms in the test phase, so that these c_n be improved and

refined continuously with the number of test_.

Preclsely concerning the last aspect it should be empha-

sized that our developmental work in pattern recognition proces-

ses for US inspectfon will not terminate in an independent system,

that has'a possibly incomprehensible or even controlled existence

_ separate from physical considerations and practical inspection _ "

experience, but rather a system that is technically highly

developed, capable and versatile, yet as simple to use as a

conventional US device, a system to serve US inspectors as a tool

for evaluating CFRP materials with all requirements according to

the best knowledge and state of the art. This also includes

detection of manufacturing defects previously undetectable

without destruction, for which in our opinion pattern recognition_ offers a-solution.

ii

5. REFERENCES /_!/

i. Rose, J.L., Y.H_ Jeong and-M.J. Avioli, "Utility of aProbability-density-function Curve and F-maps in Composite-material Inspection." Experimental Mechanics, April 1982,p. 155.

2 Rose, J.L. and G.P. Singh, "Stress Corrosion Cracking vs.Geometric Reflector Classification Analysis for 304 Aus.tenitic Stainless Steel Pipe Weld Specimens," NinthWorld Conference on Nondestructive Testing, 1979, Melbourne,Australia.

3. Rose, J.L., "A 23 Flaw Sorting Study in Ultrasonics andPattern Recognition." Materia]s Evaluation, July 1977,p. 87.

4. Carson, J.M. and J.L. Rose, "An Ultrasonic NondestructiveTest Procedure for the Early Detection of Fatigue Damage andthe Prediction of Remaining Life." Materials EvaluationApril 1980, p. 27.

5. Rose, J.L., M.J. Avioli and M. Lapides, "A Physically ModeledFeature-Based Ultrasonic System for IGSCC Classification."Materials Evaluation, December 1982, p. 1367.

6. Rose, J.L., M.J. Avioli and R. Bilgram, "A Feasibility Studyon the Nondestructive Evaluation of an Adhesively BondedMetal to Metal Bond; An Ultrasonic Pulse Echo Approach."British Journal of Nondestructive Testinq, March 1963,p. 67.

7. Rose, J.L., "A Feature Based Ultrasonic System for ReflectorClassification." In: NewProcedures in NondestructiveTesting, P. H611er, springer-Verlag, Berlin-Heidelberg,p. 239.

8. Duda, R.O. and P.E. Hart, Pattern Classification and SceneAnalysis, J. Wiley & Sons, 1973.

*************************************************************

Key to figures [pp. 15-27]

F__i.g___!.US signals, top: example of "good" signal, bottom:example of "defect" signal.a. smoothed and rectified b. time

12

Fiq. 2. Diagram of ultrasound testinga. workpiece b. defectc. coupling agent d. testing heade. (receiver only) f. (transmitter and receiver)g. screen image h. back wall echoi. defec_ echo J. surface echo

Fiq. 3. Measurement systema. ultrasound testing device b. testing headc. analog/digltal _onverter d. computere. printer f. graphics terminal /

Fiq. 4. Examples of 2 ultrasound signals from delaminationexample.a. "good range"

_ig. 5. Definition of signal features of envelope signala. envelope signal b. timec. rise time d. pulse width at %e. decay time f. area ratio .....g. centroid position of area h. asymmetryi. standard deviation of curve pointsJ. deviation from Gauss bell curve_. position of maximum _

Fiq. _. Probability density for feature values Nos. 4 and 5 indelamination examples

_. Grouping of both classes for two different featurepairings. The pairing 6/7 (below) is suitable to separate allmeasurements of both classes with a line; the upper one is not.

Fiq. _. Compilation of values of 8 sample classes with hardeningdefects. [Commas in numbers = decimal points]a. sample No. b. classc. mean ILS value d. [expansion unknown]e. quota f. finding rateg. hardening pressure h. defective vacuum sacki. hardening temperature J. temperedk. untempered 1. cooled too quicklym. separating film between layersn. s_l_cone lubricant between layerso. good p.---_ormoisture aging)q. _(for weathering) r. (for adhesion defects ...s. total t. manufacturing/defect

13

IE

i___L__.9. ,,Performancematrix" as means of representing results ofclassification. [Commas in numbers = decimal points]a. actual defect class b. found classc. total d. quotae. detection f. meang. employed features

FiG. IO. Compilation of best features/feature combinations formultidirectional laminates (8 classes) [Commas in numbers =decimal points]a. number of features b. employed featuresc. mean criterion d. total rate (accuracy)

FiG. ii. Example of 4 hyperquadratic discriminant functions for1 feature. For x4 = 1.01, e.g., class 4 is right.

Fiq. 12. Decision plane for features 12/19 for multidirectionallaminates (8 classes)

Fiq. 12. Compilation of "hit quotas" for ul.timumfeaturecombinations for the 6 types of fabric laminates. [Commas innumbers = decimal points]a. laminate type b. number of_layersc. of employed features d. classese. total criterion f. total rate (accuracy)

14

a

GEG_TTET.OR|G|NALPAG__SOF POOR QUALITY GEEICHGERICHTET

-° fIE-SIGNAL .....

Blld 1: US-S|gnale, oben BeJspiel for "Gut"-Signa]

_g. _ unten BeispJel fur "_enler"-Signal_|1277

15

\IERKSTOCK

ORIGINAL PAGE I:'OF POOR QUALITY c

KOPPELMITIEL

dPRUFl<OPF

(SENDER + f

EUPFANGER)

...-.-.--.-.

bFEHLSTELlE

I I

I I I

I I ' I.I I

II I

III

I I

(A-SCAN)

e(NUR EMPFANr-ER)

dPROFKOPF

SCHJRMBJLD g

MBJ1

r-----t. ,...--1 ,--, .,

I I'------t.-

hRUCKWANO-r_ECHO

FEHLER- i

ECHO ....OBERFlJ.CHEN- j

ECHO

t3tld 2: PrJ.nzipdarstellung der--UltraschallprUfunq.

£ig.2

~O~ln11 IL- . ..J

16

- I o

a Ultraschall-b

PrUfger_t _ kopfPr_f-• [

z_ Analog/Digital-wandlerc I Monitor

7,

!t:_,_._ .........._ d e!_ Rechner _ Drucker

_,_'

Grafik-

terminalf

. BiId 3 : MelBs_'.ster,_

Fig. 3

~ ..

.- -""-"-"""~"'--~- -._- _._~

. " ..... ";~~

o...,'.........

........, "" .•'=.;. • . . ' 1 •

k"'O urd Itlq .." ..~ 1,,"»1 .. I .. I" .,..~,.t"rt ttIausdruc1aht..,. 1tJfJfI"';t~n ]uwtttfMhar"(9hJftge''' ,,~r."tIf"Nen

IU SchftdfHWttUl1 ,lit,. 11.-.. hi,. 'ur Mra J all ",., '-at"'""""''''Q ode' v",,,,,,,, h ..JfttI~I,... r ,"118Qt~ \tf'Uhrt",nf'ft

~

.

!;11 ~r-:1fl't---· -~I\f;,--' ­r ~l~ ~

TIME

n

nii r::-LIr.t u

ni 1.Iill !P

1"rOo

rr1 G~

,..J ..,'"I:.-----..,.r -,"-._u_..;;... _

FREOUENcY

,I ~if··; t;" :~;;ti

i·: !,,· ..I ;~ 1: • t; n •• "1"1 ',.:::::.L,,_-_... i~j\r'-;=i"U 'J V

......: II

A ,

n::.:: :L

r...;:~ i: =1 :11: I 'I

n : i; ~,

i1 t :.! ~.'I ....j: L

! n! tl lI • u •

I" In~V'U ~ "-_.. ~

fP.EDUHI( f

·~ ii

I,!:l\

~ I •

d \• J :

/. \, ,

FREOUENC'f

RNRLYTIC 3PErTRU~i'i!1

.,

!\j t. ,;" ij \ /\~

~'---.;::"',-,,",,-,..----,-'--:...._-'""'!~ "Gut-BerrICh" "De1am ina t Ion 11

SlId 4:. - aeiSDlele fOr 2 Ultrascnallslgnalc des DelaminatlonsbelsPlelst,

F".....~~' ., .. ,.... ".... ,..... -.. ,-.,. ,

-'c(

f5 60 ~

100 S

-. 90 ~>~

ORIGINAL PAGE IS.PF. POOR QUALITY

I.. - ..-. ,

,,. - _. -- --,-- - ,--,-

---------_._-_.-----_.- .-

- - - .- -- - - - _. - - _.. -....... - _.. -- - .

25 %a

.-en-'-':::>

::x::

. ~..T2 T3 T4 T5 T6 T7

Z E I Tb

Fl = T3 - Tl = ANS1IEGSZ&ITc

F2 = 16 12 = PULSBRE ITE BE I 60:' % d

F3 = 17 11 = PULSBREITE BEl 25 %de ".

Fit = T7 - T5 = ABKLINGZEIT T".a/~ l;...,

F5 = FUi.CHENVERH.6.LTNI ~ = JV.UIfYell:o ~/.t

F6 = SCHWERPUNKTSLAGE DER FLACHE g

F7 :: S, A~:,: hRJf.,E~::'1 CHW~G DER KURVENPUt':KTE i

Fg = SKEWNESS (ASYMMETRIE)h

Fg = KURTOSIS (ABWEICHUNG VON GAUSS-GLOCKENKURVE) j

FIO= Tit = LAGE DES MAXIMUMSk

-B·; 1d- ~(---Defi ni.tio n von S i 9na1mer kma 1en .des HU11 s i 9 na1s

Fig. 5 (Video Envelope (V))

. :'~1' L- '

19

I'J ~.r:1£.. --

ORIGINAL FrAaS ftOF POOR n~ 11\1.rrv

I1. 79

; • (j~ -;.. Ut

/-;.~\/1

II \/ . '\II '\,# \f \.II ..

Ii \// \\n v

I' ~

II \'.~~<~-

Gl~-I---""':::::::"""--==r=~~~~==bF===~=-----,

6 0.23 -O.Go J.03 0.23 ~Li!C frl.56 0.72FEATURE ~ 014

ttl

(f)swe>,d.CC N:::JUs

I GLL"(.D _

0..-10-

sc:!:)

·ro

• Q-.

!tp Q

li·~ ·t -! ~. 20 {[I.8liloC" :.\ \ ~ ~• <.q·:t -_•• C;;,I

~I ~ .. ~ s~ ~t~: ...... ·~ i -= l

S.: ~ '( ::;r-

.~ '!!rll~ G~ ... 11 t9n.:-~ -..oJ ~ ~ I C\It., i~ l"")2)'

H;f m srL~.:. _ ... Co W s.,';.;: - .';;>::;14

CCC\l• ~

UsI lSI

LL~0-'0-

ss

•co

tli 1d 6: . Wahrsche 1nIi chke i tsd i ctlte-fur Nerkma 1swerteFig. 6 Nr. 4und Nr.-5 'be i Delarn-i nat i onsbe isp i e) e

'.11777 ~o....- .J20

r...,_

_o_ 2

, I_ ORIGINALPAG_ ISOF POOR QUALITY

t_ 2 lU_'_. 2

I

q-

" I I I I I

"._'-" el.00 0.20 0._10 0.60 0.80 I._0•_{i: FEAI31ZI5

. clZ

•'.' " -" 2

2 2

. _..1i /

f r._.- 2 /

el: 2 @ 2_,,_ ./__i--

• f_._l

2/ I IrJ_ / I l

/ i/

{9I I I I --3

_, 00 0..20 0, _0 0.60 0.80 I •00FE.R_IZ)6

Bild 7: Gruppierung der beiden Klassen fur zwei verschiedeneMerkmals-Paarungen. Die Paarung 6/7-(unten) ist ge-

Fig. 7 vigu_c, alIe Messun_en der beiden Klassen mlt einerLinie zu trennen, dle-obere nicht.

21

•,i,"

i- f~ t: i~ -•. j'.:

,;

"j,

'-' ~;.:; i~

~ f'~ j

~.. •

tP

... "" ........ ..... -... - . ... ~, . ,

Hr. b ILS-Wert d rateua--

N/1IVI12 ~ ~

C21 I 50 ~ Aushartedruckg 2 62.4 80,0 93,3-C21 I I 2 59.8

C22 I defekter Vakuumsack h 3 48.2 90,0 100,0C22 II . 3 48.5-

C23 I Hartetemperatur 150°, ge- 4 67.2 100,0 100,0C23 II tempert i,j -4

C24 I HKrtetemperatur 150°. un- 5 66.1 80,0 90,0C24 11 getempert i,k 5

C25 I AbkUhlung zu schnell 6 50.8 70,0 100,0--C25 II 1 6 :>0.3

C26 I T fl' .. 2./3. Lage 7 66.6 80,0 : 100.~C26 II renn 0 le .w. D /9 Lage 7 5:--.9 d£::.'m u..i

C27 I S'l'k f tt 2./3. Lage 8 62.3 90,0 ,C27 II 1 1 ~n e ZW· 8./9. -Lage 8 46.9 li~ 1. 100-,0

C31 ,

C32,

gut (fUr Feuchtealterung) 1 73.0 IC33 ~ --C34 O,p

C35 IC35 11 >

C36 I gut (fUr Klimaalte'fung) 1 72.8

C36 IIo,q

- >- 87,5 -C41 IC41 I I gut (fUr Klebfehler I) 1 71,0C41 I II

'C41 IV o,rI I

C42 1

J..

C42 II gut (fUr Klebfehler JI) 1 71.3C42 IIIC42 IV o,r

r ._-- -~ Summe:5

84,7 97,6

Fl.g. 8

-B+1-d8:Zusammenstellung der Werte der 8 Probenklassen mit Ausharte-

fehlernORiGiNAL r-l.. :':': :~

OF POOR QUALITY

'\ :'7 l- --.;:....-. _

22

r

tats_l lche Fehlerklasse: a

ORIGINALPAGE IS

[OIE.POORQUALITY 1 2 3 4 5 6 7 8

I 15 I 0 " 0 0 0 0 0

2 I 9 O. 3 1 3 1 0

_. 3 0 0 9 0 0 I 1 1

b 4 0 0 0 5 4 1 0 0

}: u:,

5 0 0 0 1 3 5 2 0o r

;_ v.'._ .

._ 6 0 0 0 0 0 0 0 1;-"_= C(_ 7I " 0 0 0 0 "0 _ 0 6 1

_i ! '_8 0 0 1 1 2 0 0 7

Summe:c t6 10 10 10 10 10 ' 10 10

Quote:d % 60,5 93,.7 90,0 90,0 50,0 30,0 0,0 60,0 70_0.

Detekt.e:%98,6 90,0 100,0 100,0 100,0 100.0 100,0100,0

OJoteal: 9%7g

vr, r'l_:_tt_f ,', M_rlzt_lim, _ 1 0 1 0f

MI't : 96,2 .....

Blld 9: "Performance-Matrix" als Darstellungswelse forFig. 9 Kl.assl f 1kat lonsergebnlsse

Q.|/_77 , m

23

Zahl benutzte NI ttel_ert- GesaMt-

der RerkmaIe kr ! ter !um quoteb c d

_rkmalea

5 87,_% 37,7 %

8 82,8 32,7

17 87, 1 36, 4

0 I 6 86,3 35,2

16 86,3 35,2

= . o •

!,- -"=_ 12 19 96,2 60.5. €.

" -" "_ 92, 3 52, 2_.- 5 14_. -

O12 19 25 96.9 63,0

5 8 ]4 96,2 60,5

5 11 14 96,2 71,7

3 6 19 24 96,2 63,0

12 19 24 .... 96,2 65,5

. o .

t • •

Bild 1-0: _Zusammenstellungder besten Merkmale/Merkmalskomb|nationen,,_i _,u,,_._,.._;e LaT,t_,ate(8 Klassen)Fig. lo for'" ..........

"'1 i

// I .,,,o.//_f,,..=.,0_ o_ ' "_ \_ "_-

,i

----- /i'l t/F#T " ' "I .._*-° • _.t

' " " ' Ii ORIGINAL I, t J PAGE !$ ", ',

I / # / , ',

, ,, ! j OF POOR-QU'J,LITY , -,/ / # • (,,

7 € if * % '_# t i ' \ '_i " 'i \ ",i! .P".. I •

VI I/ I I ' \l" r , i \

,# t I % !

# / Iu I I% q

q/ 4 %I---- # I" I \ II%1%%I !

' ' ! I-- ,,# # - \,,

:'::- '-'- 'J " I:-_--_ '- t)4 I I ",

-'- I" :# " " I'-r- #_ '_i

-_'=F. -- , I %'1

_:l _ I l "'-'IIi[ t,I -1_1_ i I _' l II

' I Ill

" I r" _ _ '' '"' I "" "- "-" 'lj1 l"'ll I11 i "_1 I11 k I I IiI ka li.J i _i, _L_J .i. • _ I

i ItI Id

; I!

" II! "tIII . II - ° II

4 i"*'.l II. ,--1 I

xq = 1,01'-_ 7_''."J f-_'-'- E '__ _U, _- Ll-'t i l.J° ,I ". _ "_

Blld 11: 8etsplel fur 4 hyperquadratlsche Diskrimlnanten-

F±_. _._ funktlonen f_Jr 1 Nerkmalfar x., -- 1.01 1st z, B, Klasse n rlchtig

'_° "7

25

0.1921£+00

777777717777777

:77~777777777777

777777777777g77777777777777777777777 77 777 777 77777777777777777777777777777777

o i

0.1,9S£+01

BlId 12:Fig. 12

Entscheldungsebene fOr merkmale 12/19 fOrnultidire<tiorBle L~~ir:ate (8 Klassen)

ORIGINAL-PAGe ISOF.. POOR QUAliTY..

1/777 '------------------------------------- --J26

,.

?

tYP a zah~ der benutzten Merkmale c.. __..

6 76~ 3 86~3 86~3 97~ 5~--

ACl 12 9L7 96~9 99~O 100~O

(5 Klassen) 18 ]8~~8 92~5 98~8 100~O

7 85~0 98~3 100~0 -B 13 91 ~ 7 100~O 100~0 -d

(4 Klassen) 19 93~3 100~O - --

e Sl.J1JJeli<r iter 1L1T1 (%)

~-.I

,.....

". j

~~O6 621 0 74~O 88~0.A d 12 8L7 91, 7 98~3 1(010,

(5 Klassen) 18 74~O 86~O ~.,O 1CDIO:

.. '

.

7 82~5 97~5 1010 -.~

B 13 67~5 95~O lOJ~O -d~~

(4 Klassen) 19 77~5 la.'~O - -

f Gesamtquote (%),r_....'

Bild 13: Zusarnrnenstellungder Trefferquoten der oP.t1 rna Ien Merkma 1s-Fig. 13

~omblnatloneRfOr~le 6 Gewebelarnlnattypen

I ,

27

----~----~--,~-----